近年來遙測技術在觀測地球的各方面應用越來越普及，衛星觀測可快速獲得大範圍的地表資訊。針對崩塌災害問題，由颱風或強降雨造成的崩塌往往伴隨著多雲的大氣狀況，可能限制了光學影像的觀測能力，使得較不受雲霧限制的雷達觀測資料具有較大的優勢以提供即時提供災害資訊。為分析雷達資料的應用性，本研究使用合成孔徑雷達資料分析兩場崩塌事件中的崩塌地後向散射變化，分別為 2015 年因蘇迪勒颱風在台灣北部強降雨造成的崩塌地，以及2018年北海道大地震造成的丘陵地大面積崩塌。本研究使用歐洲太空總署之 Sentinel-1 C band 免費雷達資料，為了偵測崩塌地並分析崩塌地之後向散射變化，本研究發展了一個影像融合方法來減低地形對雷達回波造成之影響；另外，不同於使用單一事件後影像、或進行事件前後的影像變遷分析，本研究疊加多時期之雷達影像進行崩塌地的判釋。為此本研究蒐集災害發生前三個時期的影像 (T1-T3) 與災害發生後三個時期的影像 (T4-T6) ，共六個時期之影像，並設計四種實驗，應用監督分類器-支援向量機 (Support Vector Machine, SVM) 對上述兩場崩塌事件進行崩塌判釋。本研究設計了四種不同的多時序影像分析實驗，來檢視其崩塌判釋之正確性，四種實驗分別為：(1) 僅用災害發生後的一張影像 (T4)；(2) 應用變遷分析概念，採用最接近災害發生前的一張影像 (T3) 與最接近災害發生後的一張影像 (T4) 進行分析；(3) 試驗即時監測之可行性，採用四個時期的影像，分別為災害發生前的 T1, T2 和 T3，以及災害發生後最近期 (T4) 之影像；(4) 採用全部六個時期的影像，利用較多的時序資料進行分析。同時，本研究在 SVM 分類過程也將三種極化資料的使用分別進行試驗：(1)僅使用VV 極化資料、(2)僅使用 VH 極化資料以及(3)同時使用VV及VH 兩種極化資料。 研究結果顯示，使用融合影像相較於用單一觀測模式之影像能較有效地偵測崩塌地的回波強度變化。其次，在利用多時序雷達資料進行影像分類時，四種模式中，以模式四 (T1-T6) 使用 VH+VV 融合資料有最佳分類成果。但若考慮即時、快速判釋的可能性時，則模式三 (T1-T4) 配合 VH+VV 資料在兩場事件針對 0.5 公頃以上的崩塌地，可以達到 75% 以上的正確率。而本研究之兩場事件的主要分類誤差為對崩塌地的過度分類。最後，本計畫期望此研究成果能有助於精進現有之山區崩塌災害監測方法，對崩塌災害管理與減災措施等能有具體貢獻。 ;The Earth Observation Satellites (EOS) are providing unprecedented opportunities to detect changes and assess economic impacts in case of disasters. To detect landslide hazards for a wide region, remotely sensed data has been widely applied due to its efficiency and low cost. However, the cloudy condition during a typhoon may limit the application of optical data. For an emergent monitoring task, Synthetic Aperture Radar (SAR) is a suitable tool for detecting landslides in cloudy and rainy weather. In this study, multi-temporal SAR images were analyzed to identify the backscattering changes over landslide surfaces and incorporate them into a supervised image classifier. Data from two landslide events, Typhoon Soudelor (2015) in northern Taiwan and Hokkaido earthquake (2018) in Japan, were collected to perform the landslide mapping tasks. The Sentinel-1 C-band SAR images from two observing modes (ascending and descending) were fused to reduce shadow effects in the mountainous region. For analyzing the characteristics of multi-temporal SAR backscattering of landslide, this study collected image datasets, including VH and VV polarization, from six acquisition time: three are from pre-event and other three are from post-event (denoted by T1-T6, respectively). Four experiments are designed to perform the support vector machine (SVM) classification. Experiment 1 only uses images from T4 (backscattering dependence classification). Experiment 2 uses images from T3 and T4 (change detection). Experiment 3 uses images from T1-T4 (rapid mapping capability). Experiment 4 combines all of the images from T1-T6 (full time-series dataset). The results show that Exp. 3 and 4 can produce better results than the other experiments, especially when VH and VV polarization datasets were all incorporated in the classification scheme and Exp. 3 could be potentially applied for the urgent mapping of landslide hazards. The main mapping error is due to the misclassification between vegetation and landslide, and the proposed method is much effective in detecting landslides >0.5 ha with mapping accuracy >75%. Overall, the experiment results indicate the usefulness of the proposed method in detecting event-landslides using multi-temporal SAR images.
